Failure in Training due to Empty Validation Metrics Dictionary when run_validation_freq is greater than 1

[AlfredQin]

🐛 Describe the bug

Bug: Failure in Training due to Empty Validation Metrics Dictionary

Description

When the run_validation_freq parameter in training_params of Trainer is set to a value greater than 1 (e.g., run_validation_freq: 5), the training process fails. This appears to be due to an issue with the validation metrics dictionary (valid_metrics_dict).

Details

The valid_metrics_dict is initialized as an empty dictionary. However, if the condition (epoch + 1) < run_validation_freq is met, the dictionary remains empty and is passed to the _write_to_disk_operations function. This results in an error when attempting to access validation_results_dict[self.metric_to_watch] in the _save_checkpoint function because the validation_results_dict is empty.

Here is the problematic code:

# RUN TEST ON VALIDATION SET EVERY self.run_validation_freq EPOCHS
valid_metrics_dict = {}
if (epoch + 1) % self.run_validation_freq == 0:
    ...
    valid_metrics_dict = self._validate_epoch(context=context, silent_mode=silent_mode)
    ...
...
self._write_to_disk_operations(
    train_metrics_dict=train_metrics_dict,
    validation_results_dict=valid_metrics_dict,
    ...
)
...
...
def _write_to_disk_operations(
        self,
        train_metrics_dict: dict,
        validation_results_dict: dict,
        test_metrics_dict: dict,
        lr_dict: dict,
        inf_time: float,
        epoch: int,
        context: PhaseContext,
    ):
     ...
        # SAVE THE CHECKPOINT
        if self.training_params.save_model:
            self._save_checkpoint(self.optimizer, epoch + 1, validation_results_dict, context)
...
def _save_checkpoint(
    self,
    optimizer: torch.optim.Optimizer = None,
    epoch: int = None,
    validation_results_dict: Optional[Dict[str, float]] = None,
    context: PhaseContext = None,
) -> None:
    ...
    metric = validation_results_dict[self.metric_to_watch]


### Versions

Set run_validation_freq > 1 in the training parameters.
Run the training process.
Observe the error when the training tries to save a checkpoint.

[BloodAxe]

Thanks for the bug report. We will investigate this behavior

[hakuryuu96]

@BloodAxe please, see this PR. I've attempted to fix this.